colin barre-brisebois - gdc 2011 - approximating translucency for a fast, cheap and convincing...

Approximating Translucencyfor a Fast, Cheap and Convincing

Subsurface Scattering Look

Colin Barré-Brisebois (speaker)Marc Bouchard

Game Developers Conference 2011

Agenda• Prelude – Real-Time Demo• 1. Translucency in Computer Graphics• 2. Technique Details• 3. Implementation Details• Q & A

Fig. 1 – Real-Time Translucency in Frostbite 2

Real-Time Translucency Demo

Translucency in Computer Graphics

The quality of allowing light to pass partially and diffusely inside media.

Translucency

Fig. 2 – Translucency in Athena Statue (left) [BarréBrisebois11], and child’s hand (right)

• We rely heavily on BRDFs for describing local reflections

– Simple and effective for opaque objects• However, many objects in nature are (partly)

translucent– Light transport also happens within the surface– BRDFs are not sufficient

• BSSRDFs allow for an even better simulation – But are usually more/too expensive

• In our case, we chose BSDFs (BRDF + BTDF), with some elements of BSSRDF

Translucency in Computer Graphics

Fig. 3 – BRDF, BTDF and BSSRDF

Real-time translucency and derivatives come in different flavors:• The more complex, but (relatively) expensive

– [Chang08] Texture-space Importance Sampling– [Hable09] Texture-space Diffusion blurs, for skin / SSS– [Ki09] Shadowmap-based Translucency / SSS

The State of Translucency

Fig. 4 – Texture-space Importance Sampling [Chang08]

Fig. 5 – Texture-space Diffusion [Hable09]

Real-time translucency and derivatives come in different flavors:• The simpler, but faster

– [Sousa08] double-sided lighting & attenuation for foliage

• We want: fast like [Sousa08], and nice / complex results like the others

The State of Translucency (cont.)

Fig. 6 – Foliage Translucency from Crytek’s Crysis

And we got...

Fig. 7 – Real-Time Translucency in EA DICE’s Frostbite 2 Engine

And we got... (cont.)

Fig. 8 – Real-Time Translucency (Skin/Left, Hebe/Right) in EA DICE’s Frostbite 2 Engine

Technique Details

Overview• We don’t want to rely on additional depth maps and texture-space blurs

– Requires more memory (i.e. for depth maps)– Requires significant computation (i.e. for texture blurs)– The previous are still feasible, but what if we could do without...

• In essence, for convincing translucency, the light traveling inside the shape:– Has to be influenced by the varying thickness of the object– Has to show some view & light-dependent diffusion/attenuation

• We only really need a simple representation of inner surface diffusion– Most users will be convinced even if not fully accurate! – Also, if the effect is cheap, we are free to use it everywhere!*

Overview (cont.)

Fig. 9 – Direct and Translucency Lighting Vectors

Local Thickness

Fig. 10 – Local Thickness on Hebe

Computing Local Thickness• We rely on ambient occlusion for computing this info:

1. Invert the surface normals2. Render ambient occlusion3. Invert the colors and store in texture

• Can also be stored in vertex color, if tesselation allows• Similar to [Sousa08], but streamlined for meshes of varying

shapes and sizes.

Computing Local Thickness (cont.)

Fig. 11 – Local Thickness Texture for Hebe

What About Subsurface Scattering?Even if not mathematically perfect,our technique gives an impressionof SSS:• Local thickness approximates light

transport inside the shape– Different color for direct and indirect light

gives convicing light color subsurface transfer

• View-oriented distortion and attenuationgives an organic result, breaking theuniformity

Fig. 13 – Our Technique, combined with Skin Shading

Implementation Details

half3vLTLight=vLight+vNormal*fLTDistortion;halffLTDot=pow(saturate(dot(vEye,‐vLTLight)),iLTPower)*fLTScale;half3fLT=fLightAttenuation*(fLTDot+fLTAmbient)*fLTThickness;outColor.rgb+=cDiffuseAlbedo*cLightDiffuse*fLT;

• Generates approx. 13 ALU instructions (based on platform)– More performance details in following

slides

• Can precompute powers in order toget rid of the pow()

Code

Fig. 9 – Direct and Translucency Lighting Vectors

Managing the Data at RuntimeAt runtime, we have several parameters to manage, some per-light, some per-material:• Per-light parameters are used during the deferred light pass• Per-material parameters are stored in the g-buffer

– Can also be stored in a separate buffer, if space-limited– Some parameters make more sense per-light, some per-material– This is very specific to your g-buffer setup– Might require some clever packing/unpacking

• For packing more parameters on a per-material basis, instead of per-light

– Using this technique with deferred shading definitely brings-the-thunder!

Managing the Data at Runtime (cont.)half3vLTLight=vLight+vNormal*fLTDistortion;halffLTDot=pow(saturate(dot(vEye,‐vLTLight)),iLTPower)*fLTScale;half3fLT=fLightAttenuation*(fLTDot+fLTAmbient)*fLTThickness;outColor.rgb+=cDiffuseAlbedo*cLightDiffuse*fLT;

fLTAmbient• Ambient value• Visible from all angles• Representing both front and back

translucency that is always present• Optimally, per-material

Fig. 14 – Ambient Term

Managing the Data at Runtime (cont.)half3vLTLight=vLight+vNormal*fLTDistortion;halffLTDot=pow(saturate(dot(vEye,‐vLTLight)),iLTPower)*fLTScale;half3fLT=fLightAttenuation*(fLTDot+fLTAmbient)*fLTThickness;outColor.rgb+=cDiffuseAlbedo*cLightDiffuse*fLT;

iLTPower• Power value for direct translucency• Breaks continuity, view-dependent • Can be optimized with pre-computed

powers• Optimally, per-material

Fig. 16 – Power (4/Left, 12/Right) Term

Managing the Data at Runtime (cont.)half3vLTLight=vLight+vNormal*fLTDistortion;halffLTDot=pow(saturate(dot(vEye,‐vLTLight)),iLTPower)*fLTScale;half3fLT=fLightAttenuation*(fLTDot+fLTAmbient)*fLTThickness;outColor.rgb+=cDiffuseAlbedo*cLightDiffuse*fLT;

fLTDistortion• Subsurface Distortion• Shifts the surface normal• Breaks continuity, view-dependent

Allows for more organic, Fresnel-like• Optimally, per-material

Fig. 17 – Distortion (None/Left, 0.2/Right) Term

Managing the Data at Runtime (cont.)half3vLTLight=vLight+vNormal*fLTDistortion;halffLTDot=pow(saturate(dot(vEye,‐vLTLight)),iLTPower)*fLTScale;half3fLT=fLightAttenuation*(fLTDot+fLTAmbient)*fLTThickness;outColor.rgb+=cDiffuseAlbedo*cLightDiffuse*fLT;

fLTThickness• Pre-computed Local Thickness Map• Used for both direct and indirect

translucency• Attenuates the computation where

surface thickness varies• Defined per-material

Fig. 18 – Local Thickness

Managing the Data at Runtime (cont.)half3vLTLight=vLight+vNormal*fLTDistortion;halffLTDot=pow(saturate(dot(vEye,‐vLTLight)),iLTPower)*fLTScale;half3fLT=fLightAttenuation*(fLTDot+fLTAmbient)*fLTThickness;outColor.rgb+=cDiffuseAlbedo*cLightDiffuse*fLT;

fLTScale• Scale value• Direct / Back translucency• View-oriented• Should be defined per-light. This

makes it the central control point

Fig. 15 – Scale (1/Left, 5/Right) Term

All-Together

Fig. 19 – The Final Result on Hebe

Minimally, translucency can be stored in the g-buffer as asingle greyscale value:

Based on your game, this can be enough. The color will then only originatefrom the light sources (and also diffuse albedo).

Deferred Shading G-Buffer Setup

Fig. 20 – Our G-Buffer, with greyscale Translucency

Deferred Shading G-Buffer Setup (cont.)

All objects here are relying on a greyscale value for translucency

Fig. 1 – Real-Time Translucency in Frostbite 2

Better results will be achieved if translucency is a color (here,with some packing):

This translucency color, representing our inner surface color diffusion, willbe combined to the light color and the material’s diffuse albedo.

Deferred Shading G-Buffer Setup (cont.)

Fig. 21 – Our G-Buffer, with coloured Translucency (packed/offset)

Green LightWhite AlbedoRed Translucency -># # # # #

# # # ## # # ##

Deferred Shading G-Buffer Setup (cont.)

Fig. 22 – Real-Time Translucency in Frostbite 2

<- Blue LightWhite Albedo &Translucency

PerformanceXBOX360 PLAYSTATION3 PC(DX11)

Full‐ScreenCoverage 0.6ms 1.0ms 0.03ms

Instruc9ons 13 17 12

• PS3: Takes advantage of our light-tile rendering on SPUs–See Christina Coffin’s ”SPU Deferred Shading for BF3 on PS3”

• DX11: Supported in our Compute Shader solution–See Johan Andersson’s “DX11 rendering in Battlefield 3” talk for more

DX11-related details

• PC: AMD Radeon 6970

Caveats• Doesn’t take all concavities into account• Technique is optimal for convex hulls• Doesn’t work with morphing/animated

objects– Alternative: Though camera dependent, several

cases could work with a real-time thickness approximation [Oat08]

– Alternative: Could also use an hybrid dynamic AO computation, with inverted normals

Fig. 23 – Concave Hull

Summary and Q&A

• We presented an artist-friendly, fast and scalable real-time approximation of light transport in translucent homogenous media:–Improves games visuals by adding a new dimension, with light traveling

inside shapes–Has a scalable and reasonable impact on runtime–Provides convincing results even if not mathematically

perfect

• This technique is also published in GPU Pro 2, released this week at GDC. Check out the book!

• Sergei Savchenko• Johan Andersson (@repi)• Christina Coffin (@christinacoffin)• Halldor Fannar• Joakim Svärling• Stephen Hill (@self_shadow)• Frederic O’Reilly• John White• Wessam Bahnassi• Carsten Dachsbacher• Daniel Collin (@daniel_collin)• Torbjörn Malmer• Kenny Magnusson• Dominik Bauset• Sandra Jensen

Special Thanks

• Mohannad Al-Khatib (@psychodesigns)• Colin Boswell (@bozz)

References[BarréBrisebois11] Barré-Brisebois, Colin and Bouchard, Marc.”Real-Time Approximation of Light Transport

in Translucent Homogenous Media”, GPU Pro 2, Wolfgang Engel, Ed. Charles River Media, 2011.

[Chang08] Chang, Chih-Wen, Lin, Wen-Chieh, Ho, Tan-Chi, Huang, Tsung-Shian and Chuang, Jung-Hong. “Real-Time Translucent Rendering Using GPU-based Texture Space Importance Sampling,” Computer Graphics Forum (Eurographics 2008), Vol. 27, No. 2, 2008, pp 517-526.

[Hable09] Hable, John, Borshukov, George and Hejl, Jim. “Fast Skin Shading,” ShaderX7: Advanced Rendering Techniques, Wolfgang Engel, Ed., Charles River Media, 2009: pp. 161-173.

[Ki09] Ki, Hyunwoo. “Real-time Subsurface Scattering Using Shadow Maps,” ShaderX7: Advanced Rendering Techniques, Wolfgang Engel, Ed., Charles River Media, 2009: pp. 467-478.

[Oat08] Oat, Christopher and Scheuermann, Thorsten. “Computing Per-Pixel Object Thickness in a Single Render Pass,” ShaderX6: Advanced Rendering Techniques, Wolfgang Engel, Ed., Charles River Media, 2008: pp. 57-62.

[Sousa08] Sousa, Tiago. “Vegetation Procedural Animation and Shading in Crysis,” GPU Gems 3, Hubert Nguyen, Ed., Addison-Wesley, 2008: pp. 373-385.